Policy as Data: Replay-Based Policy Dual Averaging via Advantage Regression 3YM5F3YHD5
| URI | m2://res/3YM5F3YHD5 |
|---|---|
| Kind | preprint |
| Creators | Nianli Peng, Geoffrey J. Gordon, Kianté Brantley |
| Date | 2026-10-03 |
| Subjects | arxiv cs.LG, arxiv cs.AI, arxiv math.OC |
| Scope | border |
Expressions
e1: version, preprint
| Date | 2026-10-03 |
|---|---|
| Ids | arXiv 2610.04638 |
Items in this resource
The mentions of the inventory that name this resource. Each name links to the card of its inventory row.
| Item | Kind | Where | Source |
|---|---|---|---|
| Value and advantage Lipschitzness in policy | statement | lemma F.3 | arxiv |
| The recent window tracks the current advantage | statement | lemma F.5 | arxiv |
| Annealed unregularized rate | statement | corollary F.11 | arxiv |
| Local advantage drift | statement | corollary F.4 | arxiv |
| Unregularized value bound | statement | corollary 5.2 | arxiv |
| Direct regularized windowed value transfer | statement | proposition F.7 | arxiv |
| Direct regularized optimality of the learned readout | statement | lemma F.6 | arxiv |
| Action sampling mismatch and the density weighted surrogate | statement | corollary F.12 | arxiv |
| Fitted replay value bound | statement | theorem 5.1 | arxiv |
| Performance difference lemma | statement | lemma F.1 | arxiv |
| Averaged windowed score error | statement | lemma F.9 | arxiv |
Catalogue records
- arxiv: 2610.04638 (work)